Simultaneous Detection and Conversion Among Endoscopic Enhancement Modalities

White Light Endoscopy (WLE) is the basic endoscopic imaging modality but has several limitations in enabling clinicians to detect and assess gastrointestinal diseases. More recently optical or digital imaging tools have been developed to improve the detection of subtle changes in different aspects of the mucosa: vessel, pit pattern, superficial abnormalities. These virtual chromoendoscopy (VCE) methods have proved to be superior to WLE in detecting gastric diseases and cancer precursors linked to inflammatory bowel disease (IBD). However, the different types of optical/digital chromoendoscopy are device-specific; hence, guidelines and scoring developed using one enhancement cannot be easily transferred to be used with others, cannot be applied to review past studies or endoscopic data recorded by devices from different manufacturers. The objective of this work is to develop a tool that enables endoscopists to view images that mimic the different enhancement independently of the acquisition device, possibly simultaneously on screen with baseline WLE.We propose to achieve this by first building a classifier that can detect the type of acquisition (no enhancement/WLE,VCE-IScan2, VCE-IScan3) and then apply a Cycle Consistent Adversarial Network (CycleGAN) to convert the acquired images to the missing modalities.The classifier can detect the correct acquisition type with 92% accuracy on the test set, thus using the wrong downstream CycleGANs on less than one frame every ten.